.png){kind=icon}
.png){kind=icon}
Return
A unified adaptive graph structure generation method for spatio-temporal graph forecasting
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2024.112811.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Time series forecasting has important applications in various domains of the real world. Asa method of time series forecasting, spatio-temporal graph forecasting has garnered significant attention because of its strong performance. However, traditional spatio-temporal graph models are constrained by pre-defined graph structures, which should abstract physical entities and determine their correlations. This artificial process incurs high costs and noise, which reduce accuracy. Therefore, we propose a unified method for spatio-temporal graph forecasting called Self-supervised Graph Structure Generation for Spatio-Temporal Graph Forecasting (SGX2-STGF). The use of self-supervised learning avoids high costs. We design a specific module to extract dual information from randomly initialized graph, which can effectively mitigate noise. In addition, graph contrastive learning is introduced to enhance the method's universality and performance. The proposed method can be integrated into different spatio-temporal graph forecasting models. Extensive experiments validate its versatility and superiority using three real-world datasets.
Keywords:
Time series forecasting
Spatio-temporal graph forecasting
Graph structure generation
Self-supervised learning
Graph contrastive learning